A Survey of Fundamental Frequency Estimation Techniques Used in Forensic Phonetics

Forensic phonetic investigations are routinely concerned with the analysis of recorded speech which is telephone bandwidth limited.These analyses must always entail a detailed auditoryphonetic component undertaken by a trained and experienced phonetician.[t is nowadays generally recognised that for forensic purposes.auditory-phonetic analyses must be complemented by acoustic measurements.One important aspect of speech transcription/analysis is the pitch pattem (qualitative) or fundamental frequency contour (quantitative).in the quantitative analysis of fundamental frequency.a number of devices/algorithms exist.each with relative strengths and 'weaknesses in terms of output reliability and accuracy.This paper reviews the basic operating principles of devices/algorithms found in forensic phonetics laboratories and illustrates their outputs for a high quality recording and the same speech via a standard telephone line. '2. INTRODUCTIONComparison of tape recorded voices for forensic purposes usually requires matching segmental (phonemic).suprasegmental (e.g.intonation.tempo.rhythm) and extralinguistic features of speech.The latter includes voice characteristics.such as the overall fundametttal frequency (l0) range ofan individual, which are not utilised directly to code linguistic or phonetic contrasts.The measurement of it) therefore has potential forensic signi cance with regardto suprasegmental features such as intonation and the speaker s long-term pitch range.In practice.these parameters may be important if they provide both high inter speaker differencesand low intra-speaker variance as twin criteria for speaker verification.The usefulness ofa speaker's measured f0 distribution has been shown to be: (a) a function of the duration of the sample used.and (b) context-sensitive.Bany ct al. [I] have demonstrated that aminimum speech sample duration of 90 seconds is required in order to attain a stable in distribution.It is further shown that conversational speech may be highly variable in different situations and different from read speech.which is often provided as reference material in forensic casework.Speech contexts in which in distribution differences have been noted include the acoustic situation in which the speech was produced (eg.[2]). the emotional state of the speaker (eg.[3]).and the nature of the speech material.for example.police interviews frequently exhibit a lower mean and/or modal f0 value than that for a disputed telephone recording for the same speaker and this difference has been demonstrated experimentally [4].Similarly.gender and age of the speaker [5] and tape speed at recording and playback [6] have relatively predictable effects on